Reproduce “STEP Planner: Constructing cross-hierarchical subgoal tree as an embodied long-horizon task planner” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tianxing Zhou, Zhirui Wang, Haojia Ao, Guangyan Chen, Boyang Xing, Jingwen Cheng, Yi Yang, Yufeng Yue
The Subgoal Tree Embodied Planner (STEP) framework addresses challenges in applying Large Language Models (LLMs) to long-horizon robotic tasks by constructing a hierarchical subgoal tree. It achieved a 40% success rate on the VirtualHome WAH-NL benchmark, a 34% increase over previous methods, and demonstrated superior performance on complex tasks with a real Franka Emika Panda robot.
Key insights
- Systematically addressing the contextual gap (by limiting redundant input at each decision point) and the logical gap (by breaking down abstract instructions hierarchically) significantly enhances LLM performance in long-chain robotic reasoning.
- A closed-loop framework combining subgoal decomposition with a termination model enables dynamic construction and refinement of plans, ensuring logical soundness and physical executability throughout the planning process.
- The hierarchical tree structure, specifically its ability to condense task information and reduce multi-step reasoning difficulty, is crucial for improving planning success rates in complex and long-horizon tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.